Repository navigation
feat(gritlint): gritlint lsp and gritlint export biome - #20
Merged
Merged
Conversation
Re-issued as a new dated plan (plan-discipline: a started plan is frozen). Resolves Q2 (static glibc, both targets built on each system; musl measured 80x slower) and Q10 (release wiring through pnpm-release-management at 8cd6e830). Adds R29 (cross-system byte identity) and R30 (tarball binary within 1.25x of gritlint-unwrapped), KTD32-KTD38, and open questions Q12-Q14. Applies report-only ce-doc-review findings. Solution docs point at the new plan path.
Q13 (b): KTD39 adds scripts/tools/intent-gate.sh, a changeset-intent ci.yml job and the gritlint-intent-gate flake-check case; R24 names the check. Q14: actions/download-artifact pinned by full SHA joins KTD9 for tarballs-match. Q12: no fallback approved; a static-glibc link failure or crash stops U9 with the failure.
Pin pnpm-release-management at b9b4a1d48cc1 (its main; 8cd6e830 lacks #54) in R24, KTD9, KTD35 and the release Key Decision, with the cited release.yml, changeset-check.yml and README lines re-read there. Add KTD40 (timeout-minutes: 60 on every job PR6 adds or changes; jobs that only call a reusable workflow take the called workflow's), KTD41 (the dev shell's cargo deny setup time, before and after) and KTD42 (the first-release probe with the pinned planner). Record the d96dc51 base merge and refresh the line references against the merged tree.
…s private npm/gritlint/bin/gritlint.js maps linux x64/arm64 to the binary beside it, passes arguments, stdio, exit status and terminating signals through, and exits 2 naming linux-x64 and linux-arm64 on any other platform. The manifest gains bin, files, repository and an inert publishConfig; the README's install and release sections describe the tarball and changesets. Plan: docs/plans/2026-10-10-1945-feat-gritlint-standalone-plan.md (U9, KTD34).
nix/gritlint-static.nix builds gritlint-linux-x64 and gritlint-linux-arm64 with +crt-static, each system building one target natively and the other through pkgsCross. The tarball comes from pnpm-release-management's mkPnpmWorkspacePackages at b9b4a1d over a one-member pnpm workspace, with the binaries laid into npm/gritlint/bin; the flake exports it as gritlint-tarball beside workspace-tarballs and pnpm-store, and evaluation throws if a release output would shadow one of gritlint's own packages. checks.<system>.gritlint-tarball-journey installs the tarball offline and runs gate-journey.sh against it; checks.<system>.gritlint-intent-gate pins scripts/tools/intent-gate.sh. The dev shell gains node, pnpm, the sandbox and SANDBOX_PNPM_STORE for the reusable release workflow. Plan: docs/plans/2026-10-10-1945-feat-gritlint-standalone-plan.md (U9, KTD32-KTD37, KTD39).
…systems release.yml and changeset-check.yml call pnpm-release-management's reusable workflows at b9b4a1d, configured by release.jsonc (changesets, a cargo surface for Cargo.toml and Cargo.lock, the paths gate). ci.yml uploads each system's tarball and binaries after a --rebuild, tarballs-match fails on any file aarch64-linux builds differently from x86_64-linux, and changeset-intent fails a pull request that changes crates/, apps/ or packs/ without a new .changeset/*.md. Plan: docs/plans/2026-10-10-1945-feat-gritlint-standalone-plan.md (U9, KTD36, KTD38-KTD40).
cmp -l exits non-zero for files that differ, which under pipefail ended the step after the first binary, so the second was never compared.
…e glibc rustc chooses rust-lld or GNU ld by the host it runs on, and the native and cross glibc are separate derivations whose store paths libc.a compiles in, so x86_64-linux and aarch64-linux built different bytes for both binaries (run 38135793197). Both targets now pass -fuse-ld=bfd to the C driver (rust 1.99 stabilises its own linker-selecting flags only for x86_64-unknown-linux-gnu) and link nixpkgs.legacyPackages.<target>.glibc.static, which the other host substitutes from cache.nixos.org. Plan: docs/plans/2026-10-10-1945-feat-gritlint-standalone-plan.md (U9, KTD32; R29 ruling).
…flags Temporary, reverted by the fix. Each target's build on each system writes GNU ld's link map, the cc crate's compiler commands, and the target C compiler's predefined macros and -v driver line under the flags the cc crate used for tree-sitter's core. tarballs-match prints, per target, the input sections whose size differs between the two systems' builds, and the diffs of the macros, driver lines and commands.
Cargo mixes every dependency's metadata into a unit's -C metadata, and build scripts and proc macros are host units whose hash includes the host (cargo src/compiler/build_runner/compilation_files.rs:849-856, 929-934). The same target crate therefore got different symbol names on each system. A RUSTC_WRAPPER replaces -C metadata for target units only with a hash of crate name, package name and version, crate type, target and sorted features; host units and every other argument pass through. Each crane build refuses two target units sharing a normalized hash. Removes the round-3 diagnostic (link maps, cc commands and macros, and their comparison step). The plan is superseded by a new dated plan whose KTD32 records rounds 1 to 4. Plan: docs/plans/2026-10-11-1301-feat-gritlint-standalone-plan.md (U9, KTD32).
The pinned planner gives phase=release on the first main push, owing @systemfsoftware/gritlint@0.1.0, which has no tag. Without its changelog the release refuses after pushing the tag (api-extractor-effect@8e7bd7c). The file also creates .changeset/changelogs, which version-management bump needs before it can write the next version's changelog. Plan: docs/plans/2026-10-11-1301-feat-gritlint-standalone-plan.md (U9, KTD42).
- intent-gate cases move to scripts/tools/intent-gate.test.sh, run by the flake check like mutants-plan.test.sh, with per-case ok/not ok lines - the tarball step builds with --rebuild once instead of twice - upload-artifact uses the pin mutants.yml already uses - tarballs-match drops a bash default its one step does not need - gritlint-static.nix names the C env target once and prints an absent passthrough unit with a shell default Behaviour unchanged: every x86_64-linux check builds; the intent-gate test fails its two refusal cases against an always-exit-0 gate.
…metadata wrapper Code review rulings on PR #15 (accepted findings): - tarballs-match compares the union of both systems' files, fails when either artifact is missing, empty or one-sided; upload-artifact sets if-no-files-found: error - the static build fails when the rustc wrapper normalized no target unit - KTD39 states the :(glob) pathspec the intent gate runs, and the test script that is its flake check docs(solutions): the three host-dependent inputs behind R29 (linker, cross glibc, cargo -C metadata) and the invariant that removes them. Plan: docs/plans/2026-10-11-1301-feat-gritlint-standalone-plan.md (U9, R29, KTD38, KTD39).
# Conflicts: # docs/plans/2026-10-10-1309-feat-gritlint-standalone-plan.md # docs/plans/2026-10-11-1301-feat-gritlint-standalone-plan.md # docs/plans/2026-10-11-2312-feat-gritlint-standalone-plan.md # docs/solutions/build-errors/cross-compiled-rust-binary-differs-from-native-build.md # docs/solutions/logic-errors/flake-sandbox-check-compared-two-builds-of-one-file.md # docs/solutions/performance-issues/fetch-cargo-vendor-given-whole-source-rebuilds-crane-deps.md # docs/solutions/performance-issues/gritql-target-language-debug-key-embeds-node-table.md
Root plan gate of 2026-10-12: R34/AE15/KTD47 narrow config refusals to what check does today (one shared loader, no new reason codes); R38/KTD48/AE18 specify a refused file, its batch mates and the rescan; R36/AE16/KTD51 pin one register_diagnostic call per finding on a committed two-range fixture in gritlint's engine and state what Biome-side acceptance is not proven (Q1, engine-only, no new executable); R35 adds export-unknown-rule; KTD44 states per-directory multifile batching; KTD53 states warm 8m08s and cold 9m52s against R27's 10 minutes, measured not enforced. Round-2 report-only ce-doc-review findings folded in.
PR7 (Unit B1 Done 6; plan docs/plans/2026-10-11-2312-feat-gritlint-standalone-plan.md U10, R25, R26, R31 to R38). - workspace.rs: one loader shared by check, lsp and export biome (KTD47), so a config check refuses is refused identically everywhere. - lsp.rs: Language Server Protocol over stdio via lsp-server 0.10 and lsp-types 0.97. Each event re-walks, plans, and scans only the tasks holding an open document (decide::restrict, KTD44); open documents overlay disk (Sources::overlaid). A target refusal drops that file and every directory batch holding it (decide::without, R38) and is reported on the refused file and its open batch mates. UTF-8 positions when offered, else UTF-16. - export.rs: Biome GritQL plugin for single-file css/json/js rules; refusal codes export-unknown-rule, -multifile, -language, -construct, -rewrite, -library (R35). Verified in gritlint's engine only (Q1, KTD51). - gritlint-e2e drives `gritlint lsp` over real stdio against the sandboxed binary: loadable root, refused root, and the exit-code contract. - Simplify pass: plan narrowing by value, per-document line index, one definition of the parameter-pattern shape and of each target reason code.
…ions The version PR (#18) moved the workspace members to 0.2.0 in Cargo.lock, and rustPlatform.fetchCargoVendor copies the whole lock into its fixed-output staging derivation, so its hash broke and every flake output failed on main. crane's vendorCargoDeps reads only name, version, source and checksum of third-party lock entries: registry crates come by their lock checksums and the GritQL engine by builtins.fetchGit at the locked rev. No fixed-output hash covers the vendor, and a version bump leaves its drvPath unchanged. overrideVendorGitCheckout prepares the engine checkout before crane scans it: it deletes the upstream grammar copies in resources/language-submodules, so crane keeps the copies the engine's manifests build, and the nested manifests under resources/language-metavariables, which made `cargo package --list` drop tree-sitter-php's and tree-sitter-md's sources. Docs drop the cargoDeps.hash re-pin steps and the source-git-0 path.
…rsions Records why #18 broke the vendor hash, why crane's vendorCargoDeps replaced fetchCargoVendor and not importCargoLock, the two checkout preparations the GritQL engine needs, and the version-churn gate. The 2026-10-10 lock-only vendor lesson now points at it and drops its hash re-pin step. The gritlint.nix header no longer claims a version bump keeps the dependency derivations: their names carry the version.
Review rulings for PR7 (ce-code-review, report-only): - publish handed a path's diagnostics to the first open document that resolved to it and an empty set to every other URI of the same file (a symlinked path beside the real one). Each alias now gets a copy. lsp_contract::two_uris_for_one_file_both_get_its_diagnostics fails on the previous code (timeout waiting for the second set) and passes now. - gritlint-e2e drives `gritlint export biome` against the sandboxed binary: the listing (exit 0), one exportable consumer rule (stdout equals a hand-written plugin), and two refusals (exit 2, empty stdout, one `ERROR gritlint: <ruleId>: <reasonCode>: ... Next action: ...` line). - lsp_contract pins didSave republishing the buffer, empty sets for an out-of-root path and an `untitled:` URI, and a rangeless finding at 0:0.
… into feat/lsp-biome-export # Conflicts: # docs/solutions/performance-issues/fetch-cargo-vendor-given-whole-source-rebuilds-crane-deps.md # nix/gritlint.nix
systemfsoftware-maker
added this pull request to stack #21
October 12, 2026 16:27
systemfsoftware-maker
changed the title
The hashless engine fetch made crane call builtins.fetchGit with every submodule during evaluation. CI evaluates before it restores the store cache, so each Nix job's "Every flake output" step went from about 50 s to 618 s (x86_64) and 625 s (aarch64) on run 38273376142. The engine now gets one crane outputHashes entry, keyed by its lock source string (git+<repo>?rev=R#R), so a version PR never moves it and only an engine re-pin changes engineHash. crane then fetches with fetchgit at build time, about 60 s cold, served from the store cache. engineSource is the identical fetchgit, so evaluation no longer fetches the engine at all. Two flake checks pin the crane behaviours the vendor relies on, as boundary-testing/pin-dependency-semantics requires: - gritlint-vendor-ignores-workspace-versions: vendorFor over the lock with every workspace member's version moved keeps the vendor's drvPath; given main's fetchCargoVendor vendor it fails evaluation. - gritlint-vendor-engine-copies: each locked crate marzano-language's manifest takes by path is vendored once and byte-identical to that copy; a vendor with the upstream tree-sitter-css fails it. The superseding plan records the pack and constitution comparison.
… into feat/lsp-biome-export
systemfsoftware-maker
added a commit
that referenced
this pull request
Oct 12, 2026
… refresh Review of PR #17 (cross-model peer) flagged that 'a pull request restore counts as an access' had no source. The caches API shows it: main's x86_64 entry's last_accessed_at moved to the minute #17's pull request job restored it. Symptoms also gain the live recurrence of 2026-10-12, when #19 and #20, still on the old rule, left no refs/heads/main entry.
Review run 20261012-165733-1e7bdfca (adversarial peer on MiniMax/DeepSeek): - lsp: after `shutdown` the server answers every request with InvalidRequest and waits for `exit` (exit 0) or the end of input (exit 1); a stray message no longer wedges it (#1). - lsp: two URIs of one file follow whichever buffer was edited last, not URI sort order (#2). - lsp: a failed write to the client ends the server instead of scanning on (#12). - lsp_contract pins MethodNotFound, InvalidRequest after shutdown and ServerNotInitialized after a refused initialize (#11). - export: the multifile refusal and the sibling-call refusal read the parse tree, so a commented `multifile` line or a call to a pattern a sibling rule defines is decided correctly; a rule's own helper exports (#3, #5, #9). - README: the `export biome` synopsis names `--rule` (#16).
# Conflicts: # docs/plans/2026-10-12-1740-fix-vendor-workspace-version-independent-plan.md # docs/solutions/build-errors/fixed-output-cargo-vendor-hash-covers-workspace-versions.md # docs/solutions/performance-issues/fetch-cargo-vendor-given-whole-source-rebuilds-crane-deps.md # nix/gritlint-vendor-checks.nix # nix/gritlint-vendor-engine-copies.py # nix/gritlint.nix
ryanleecode
pushed a commit
that referenced
this pull request
Oct 12, 2026
* docs(plans): save the Nix store cache on main only Pull request runs keep restoring by the nix-store-<system>- prefix, which reaches main, and stop exporting and saving. Plan only; U1 is a two-line if: change to .github/workflows/ci.yml. Reviewed report-only (ce-doc-review, 4 reviewers); findings are ruled in the hand-off, not applied. * docs(plans): refresh store-cache plan for #15 and #16 on main #15 (5e29691) and #16 (f552830) merged first. Update the renamed standalone plan path, the merged-in base, the merge-order questions, the ci.yml line references of the two conditioned steps, and the V1/V2/V4 sources, which main pushes made before this PR now supply. * ci: save the Nix store cache on main only A pull request's cache is scoped to its merge ref, so no other run could restore it, yet each one took up to 4 GB and evicted main's bundle (10.5 GB in 6 PR-only entries on 2026-10-11). #16 then restored nothing and its x86_64 job took 592 s. `Export the unsigned store paths` and `Save the Nix store cache` now also require `github.ref == 'refs/heads/main'` (plan KTD1). Every run still restores by the `nix-store-<system>-` prefix, which reaches main's newest bundle; pull requests skip the export and the save. Nothing else changes: no action, pin, key or timeout-minutes. * docs(solutions): pull request cache saves evict main's Nix store bundle Records why only main saves the store bundle: a pull request's Actions cache lives on refs/pull/N/merge, which no other run can restore, yet it counts against the 10 GB limit and evicted main's bundle. Includes the jobs API wall times before and after (#16 592 s, #17 316 s on x86_64), the alternatives rejected in the plan, and a review gate with the grep that lists every cache save step. * docs(plans): supersede the store-cache plan with a default-branch save rule The full-path review of PR #17 found the shipped `refs/heads/main` literal against CONST-T12, and the release flow's workflow_dispatch on changeset-release/main as a second non-default writer (run 38266314981). U1 had executed, so a new dated plan carries the corrected contract: R8 and KTD1 derive the default branch from the event, KTD3 and AE4/AE5 account for legacy entries on other refs, V5/V6 take the first output-changing main push after merge, and Q5 asks the operator to delete the legacy entries. * ci: derive the default branch in the store-cache save guard `Export the unsigned store paths` and `Save the Nix store cache` compare `github.ref` with `refs/heads/<default_branch>` from the event payload instead of the literal `refs/heads/main`. The restore fallback already follows the default branch by what it is; a second, literal copy of that fact could be left behind by a rename, stopping every save while every job stays green (CONST-T12). Behaviour on main is unchanged. The solution doc states the derived condition, names the release flow's dispatch as a non-pull-request writer, and corrects the claim that the cache holds default-branch generations alone while legacy entries remain. * docs(solutions): record #19/#20 evicting main's bundle and the access refresh Review of PR #17 (cross-model peer) flagged that 'a pull request restore counts as an access' had no source. The caches API shows it: main's x86_64 entry's last_accessed_at moved to the minute #17's pull request job restored it. Symptoms also gain the live recurrence of 2026-10-12, when #19 and #20, still on the old rule, left no refs/heads/main entry. * docs(solutions): state the observed restore and access times exactly The previous wording said last_accessed_at moved 'to the minute' a pull request job restored main's entry. The job logs put the restores at 16:37:36Z and 16:38:12Z and the caches API reported access at 16:39:05Z and 16:38:59Z, so the doc now states those times.
…son types Review run 20261012-194950-133db7d1 (constitution treatment, pack rules, cross-model peer): - CONST-B1/B4: `workspace` is pure (`configure` names the next read, `Configured::load` takes the read rule sources, `Workspace::select` takes the walk). `shell::load_workspace`/`shell::select` perform the reads, and `shell::Disk` implements the `lsp::Files` port the server declares. The app's composition root opens stdio and wires `shell::Disk`; no core module imports `shell`. - CONST-D1/D2: `TargetReason` (position only on a parse failure) and `ExportReason` (the six published codes) replace open `&'static str` codes. Rendered text is unchanged. - lsp: the no-progress guard returns the scan's own error; a ROOT that cannot be canonicalized serves from the given path, so a missing ROOT initializes and selects nothing, as `check` does. - Tests: a position-less `target-too-large` sibling refusal at 0:0, the missing ROOT, jsx and bare js exports; one shared test pack loader.
…d design Review run 20261012-200921-5115f566 at 0c45d15: - correctness P2: `CompiledRule::kind` reads the engine's `Problem::is_multifile`, replacing the first-line `rule_kind` test, so a `multifile` block after a comment is planned per directory and kept whole for an open file; `check`, `lsp` and `export biome` share the classifier (changeset names the `check` change). - performance P2: one source read per event serves every refusal pass. - reliability P2: a sibling rule that does not parse refuses the export (`export-construct`) instead of counting as defining nothing. - testing P2: two open documents that do not parse each get their own refusal in one event. - standards P3: `encode::message_and_next_action` is the one message join for `--format github` and the LSP (R32). - maintainability P2 / api-contract P3: the started PR7 plan is superseded by docs/plans/2026-10-12-2025-feat-gritlint-standalone-plan.md (KTD43, KTD47, KTD49, KTD50, KTD52, U10 corrected; KTD54, KTD55 new); four solution docs repointed.
…port order Review run 20261012-202737-4018f61a at 1eedb94: - adversarial peer P1 / reliability P2: one closed file the server could not read (not UTF-8) made `publish` log and send nothing, so every open document showed no diagnostics while `check` refused the tree. The server now turns a failed read below ROOT into `GritlintError::TargetUnreadable` (`target-unreadable`), drops the file with `decide::without` and reads again, so the rest of the event is published. `check`'s refusal text is unchanged. - testing P2: R35's refusal order is pinned by four fixtures, each tripping two adjacent checks; testing P3: a backslash in an exported message is pinned. - maintainability P3: `Evaluation::MultiFile` no longer stores a count derived from its own batches. Plan superseded (plan-discipline): docs/plans/2026-10-12-2059-feat-gritlint- standalone-plan.md adds `target-unreadable` to R38 and states it in KTD48 and KTD55; the living solution docs cite the new path (DEL1).
systemfsoftware-maker
marked this pull request as ready for review
October 12, 2026 22:35
ryanleecode
approved these changes
Oct 13, 2026
ryanleecode
left a comment
ryanleecode
left a comment
Contributor
There was a problem hiding this comment.
Root verdict: 5/5 green on 66a753d, 0 behind main, 0 open threads. Full-path rounds plus a delta receipt Ready to merge at this head; peers DeepSeek and MiniMax, Grok excluded. Every Round-5 item is fixed, rejected with evidence, or a follow-up with an owner (F1 Mutants, F2 Biome load check, F3 wall-clock via the repin). The 12m05s x86_64 job is diagnosed from step times (structural Cargo.lock dependency rebuild) and owned by F3. Red evidence uses only the pre-fix commit 1eedb94 or 'red not shown'; the earlier scratch-copy probes are withdrawn. Read the added tests: lsp_contract drives the real stdio server, and export_contract compares against hand-written expected plugins and runs them in the engine. The test-file clippy allows follow the repo's existing convention. Diff hunt and privacy greps clean.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Unit B1 Done 6 (PR7):
gritlint lspandgritlint export biome. Plan:docs/plans/2026-10-12-2059-feat-gritlint-standalone-plan.md(supersedes...2025..., which superseded...2312...and the PR6 plan...1301...), U10, R25, R26, R31 to R38, AE12 to AE18.#19 (
fix/vendor-lockfile-independent) has merged; this branch carriesmainthrough0e2e07b(#17) and takes #19's crane vendoring, so the new LSP crates need no vendor hash.What changes for a consumer
gritlint lsp [--config PATH] [--rules PATH]... [ROOT]serves diagnostics over the Language Server Protocol on stdio. Each open document gets one error diagnostic per finding, computed from its unsaved text, multifile rules included. The server loads whatcheckloads. If that load would refuse,initializefails withcheck's text (-32803) and the process exits 2.shutdownthenexitexits 0;exitalone, or the end of input, exits 1. Positions are UTF-8 when the client offers it, else UTF-16.target-does-not-parse,target-too-large,target-not-scanned), or a closed file below ROOT the server cannot read (target-unreadable: not UTF-8, or an I/O error such as permission denied), does not stop the server. The file gets a diagnostic with its reason code, each open document in a directory batch that read it gets one naming it, and the rest of the scan is still reported.checkstill refuses such a tree with exit 2 and unchanged text.gritlint export biome [--rule ID]prints a single-filecss/json/jsrule as a Biome GritQL plugin, or lists every enabled rule asexportableorrefused <reasonCode>: <next action>. A refusal exits 2 with one stderr line. The codes, checked in this order, areexport-unknown-rule,export-multifile,export-language,export-construct,export-rewriteandexport-library. All 15 bundled rules are multifile and list as refused. Plugins are verified in gritlint's engine only (Q1: Biome is not an approved executable).checkchanges for one rule shape (KTD54). A rule's kind now comes from the engine (Problem::is_multifile), so a rule whosemultifileblock follows a comment or a pattern definition runs per directory like every multifile rule, instead of over every file of its language at once. No bundled rule has that shape.check's loader moved intoworkspace::configure/Configured::load(pure) andshell::load_workspace(reads); its output is otherwise unchanged.minorchangeset.How a keystroke is served
decide::restrictkeeps only the single-file tasks for open documents and the multifile directory batches holding one. Open documents overlay disk throughSources::overlaid. The event's sources are read once. A read that fails on a file below ROOT refuses that file (target-unreadable), and a scan that refuses a target drops it (decide::without). Either way the rest is read or scanned again (KTD48, KTD55).Measured at
66a753donx86_64-linuxwithpackages.x86_64-linux.gritlint-unwrapped, over the PR3 corpus (Effect-TS/effect@0b17b9e2) andeffect.gritlint.json, withpackages/effect/package.jsonopen (didOpen, then didChange to publish):tsconfig.broken.jsonsiblingtsconfig.latin1.jsonsibling (target-unreadable)gritlint checkof the corpusThe host's load average was 21 to 38 during the last row; repeat sessions showed single spikes up to 296 ms on the clean tree, with no trend across 16 events.
Verification
66a753d:cargo fmt --check,cargo clippy --workspace --all-targets -- -D warningsandcargo test --workspace(16 binaries, 135 tests) exit 0. Thenix buildof all 19checks.x86_64-linux.*pluspackages.x86_64-linux.{gritlint,gritlint-unwrapped}exits 0. That includesgritlint-agent-journey, which drives the sandboxedgritlint lspover real stdio andgritlint export biome.66a753d, aarch64 included.an_unreadable_closed_sibling_is_named_on_the_open_documents_of_its_batch_until_it_reads(lsp_contract.rs:607) is red on the unmodified pre-fix commit1eedb94. I ran that commit's owngritlint-unwrappedover stdio on the test's scenario:a/x.jsonholds bytesff fe,a/package.jsonis open, and the rules arelicenseanddirectory-read. It publishes nothing for the open document. Besides its startup announce, it logs only one error:cannot read `<ROOT>/a/x.json`: stream did not contain valid UTF-8. The test's diagnostics assertion (lsp_contract.rs:623-629) expectsunlicensedat 1:2 plustarget-unreadableat 0:0, which is exactly what66a753dpublishes on the same scenario.each_refused_rule_reports_its_own_reason_code) and the R36 backslash fixture (each_exportable_rule_renders_its_committed_plugin): red not shown.export.rsis byte-identical at1eedb94and66a753d, so these fixtures pin behaviour that already held, and no pre-fix commit exists for them to fail on.757dd06,e905ecc,0c45d15anda290c70): red not shown.unlicensedplustarget-unreadablenaminga/bad.json, andshutdownthenexitexits 0.gritlint checkon the same tree exits 2 withcannot read ./a/bad.json, as before.Review (ce-code-review, report-only, five rounds)
Round 5 ran at
1eedb94(run20261012-202737-4018f61a). Lenses: correctness, testing, maintainability, reliability, performance, api-contract, project-standards (all 16 constitution laws) and learnings (41 of 41 pack rule files, 6 apply, all comply, each with a HEAD line quoted and checked). The cross-model adversarial peer was served byopenrouter/deepseek/deepseek-v4.1-flash(71 turns) andminimax-code/MiniMax-M3.1-Flash-Preview(38 turns), counted from its session transcript, and read only the diff, the file list and the evidence file. Correctness and performance returned no findings.66a753d:target-unreadablerefusal; this reverses the round-1 and round-4 "per KTD48" rejections, and the plan is superseded for it66a753d: four two-check fixtures66a753dfor unreadable targets. The rest of the log-only path is reachable over real stdio (ROOT a regular file), behaves as KTD48 specifies, and is untested: follow-up F1, ruling R0 under Round-5 residuals66a753devaluatedcount computed at two sites66a753d: derived where it is readcheckwould not select gets no diagnosticscheckis unchangedsupersedes:chain resolves the pathsexport-constructalso covers a body that does not parsecheckcompiled the body; no public code for an unreachable patha290c70: one classifier from the engine (KTD54)a290c70a290c70(plan superseded)initializeother thaninitializedexits 1initializednext0c45d15:TargetReason/ExportReason,lsp::Filesport withshell::Disk, pureworkspacetarget-too-large,jsxand barejsuntested; missing ROOT refused; loop rebuilt its error; four test pack loaders0c45d15initialize_finishreturns on disconnect or any other messageshutdownwedged the server; P1 two URIs of one file followed URI ordere905ecce905eccexport biomeCLI untested;didSave, out-of-root, rangeless finding untested; one alias of a file got nothing757dd06Final-head delta review (
1eedb94..66a753d, ce-code-review report-only, depth full)Round 5 reviewed
1eedb94; its fixes landed in66a753d. Run20261012-214525-6e25ba95reviewed exactly that delta (20 files) and returned Ready to merge at66a753d(stage logcomplete).adversarial-reviewerroute. The session transcript showsopenrouter/deepseek/deepseek-v4.1-flash(51 turns) andminimax-code/MiniMax-M3.1-Flash-Preview(13 turns) served it, with no Grok or xai. It finished in 915 s of a 1210 s deadline and read only the brief, the diff, the evidence file and the reviewed tree. It filed 2 findings.target-unreadable(peer)lsp.rs:430-433puts the read error's own text in the message, and KTD48 (plan:257) specifies exactly this conversion\n(peer, pre-existing)export.rshas the same md5 at1eedb94and66a753d, and R36 (plan:120) specifies the escape Biome unescapessupersedes:names its own former path (api-contract, anchor 50)Round-5 fixes on the final head:
target-unreadable: correctness filed nothing, and the peer checked the refusal's scope by hand. The regression test is red on the unmodified pre-fix commit1eedb94(see Verification).export.rsis unchanged since1eedb94.evaluatedfield is gone; CONST-D4 complies.No new defect survived. The three gaps the delta report recorded (two unreadable files in one event, an unreadable file seen only by single-file rules, non-encoding read errors) are ruled D1 to D3 under Round-5 residuals, each with a stdio probe.
Report errata: the rendered
report.mdprose says the peer "confirmed findings are design decisions"; in fact the peer filed them and the validator rejected them. It also names aTargetReason::NotParsedvariant (the code hasDoesNotParse) and lists 3 of the 5 packs whose rules apply (package-topology and schema-laws also apply). The verdict, counts and verdicts above come from the run's JSON artifacts.ci-wall-clock: read by hand from systemfsoftware
mainat763fd663, because gritlint is not enrolled..compound-engineering/config.yamlat66a753d(identical tomain) pins the compound packs at02bf9653and declares five packs.02bf9653predates ci-wall-clock, which systemfsoftware#742 added (merge8d071180). ce-code-review therefore loaded only the five declared packs (41 rule files), and no receipt above covers ci-wall-clock. I read its 12 rules at763fd6639979ec955dcb004e6d4d97e1cb49f9eb(#746's merge, systemfsoftwaremainwhen the delta review ran) and matched eachapplies_whenagainstmain...66a753d. Enrolment PR: none exists yet. No gritlint PR or branch has changed.compound-engineering/config.yamlsince #7 (searched 2026-10-13). Enrolment is the ci-wall-clock repin queued after this PR, which this PR does not start. The precedent at the same pin is systemfsoftware/younna-ai#38, which pins763fd663and enrols ci-wall-clock.66a753dpr-gate-runs-only-what-the-diff-affects(applies_whenline 10: a diff that changes a CI-facing flake output)checks.<system>.gritlint-agent-journeynow drivesapps/gritlint-e2e/src/lsp.rs:18(.arg("lsp")), andgritlint-nextestgainslsp_contract/export_contract.github/workflows/ci.yml:64restores (actions/cache/restore) before:78-79builds, and the diff changes no workflow. Fits in 10 minutes: measured over. x86_64 took 12m05s; see CI wall clock (follow-up F3)flake-outputs-build-once-on-main(line 5: changing a check a workflow builds)ci.yml:64), and only the default branch exports and saves (ci.yml:96,:107). This PR's job builds the dependency derivations without saving because its diff changes their inputs (Cargo.lock), the input-change case of the two the rule allows (the other is a missed restore ofmain's key)no-lane-grades-the-repos-own-docs(line 4: adding a test)crates/gritlint_core/tests/lsp_contract.rs:53runslsp::serve, andapps/gritlint-e2e/src/export.rs:74runsexport biome --rules rulessetup-under-a-minute,gates-share-one-job,no-dev-shell-on-verdict-lanes,one-stuck-job-backstop,light-checks-run-the-pinned-tool,heavy-suites-judge-every-main-commit,heavy-lanes-state-their-daily-minutes,red-main-reverts-its-culprit-same-day,release-promotes-the-staging-proven-buildgit diff --name-only main...66a753d -- .github .compound-engineering nix flake.nix flake.lock scripts .cargois empty. The diff adds or changes no job, step, trigger, timeout, heavy lane, release, light check or dev shell. For reference, run 38307881354's x86_64 job reached its first verdict step (An unsupported system is refused) 8 s after it startedRound-5 residuals: every item ruled
Round 5 (run
20261012-202737-4018f61a, at1eedb94) dispatched nine reviewers. All nine returned; none timed out. It then ended at dispatch, superseded by66a753d, so its residual risks and testing gaps were never merged or ruled. Each one is ruled below. A ruling is one of three kinds:Probes ran at
66a753dagainstpackages.x86_64-linux.gritlint-unwrappedover real stdio, with scratch trees and the fixture rulestests/fixtures/lsp/rules/{license,directory-read}.md. Test gaps are stated by reading the existing tests and the product path they drive. Earlier scratch-copy red probes were hand-applied mutants; they are withdrawn as evidence and are not cited here. No probe, scratch test or fixture is committed. Line numbers are at66a753d.Follow-ups:
main's first Mutants run over the feat(gritlint): gritlint lsp and gritlint export biome #20 merge, as test(core): kill the 72 missed mutants of full run 38092194120 #14 and test(core): only the first heading leaves a rule's message #16 were (plan R21). Trigger: the first CI Mutants report onmainthat includes the feat(gritlint): gritlint lsp and gritlint export biome #20 merge. F1 cites that report's mutant IDs forlsp.rsanderror.rs. This PR's in-diff list below includescrates/gritlint_core/src/lsp.rs:528:9: replace Server<'_>::log with (); its outcome is not predicted here. F1 adds the twolsp_contractcases the R0 and R35 readings show missing, both through the realshell::Disk. In the first, ROOT is a regular file: one ERRORwindow/logMessageand nopublishDiagnostics. In the second, a rule calls GritQLlog()and the open document getstarget-not-scanned. F1 also widensnpm/gritlint/README.md:111-112, which still says only "cannot read as UTF-8 text" fortarget-unreadable.66a753d, since unreadable targets are now refused. The body's earlier "reachable only through aFilesdouble" was wrong: with ROOT a regular file,initializesucceeds and each event sends one ERROR log ("path<ROOT>is not a usable relative path: the path is empty") and publishes nothing.checkexits 2 with the same text. That matches KTD48 (plan:257). No test observes this path. Thelsp_contractharness drops every notification other thanpublishDiagnostics(lsp_contract.rs:93-118), and no file undercrates/gritlint_core/testsorapps/gritlint-e2e/srcmentionswindow/logMessage,LogMessageorMessageType, so nothing readsServer::log's output (lsp.rs:527-532). Red not showncheckchanges for a rule whosemultifileblock follows a comment.changeset/gritlint-lsp-biome-export.md:9and What changes state it, andtests/scan_contract.rs:814pins it. Each of the 15 bundled rules hasmultifileas its first code line, so none changescheck(table above). KTD44 and KTD55 choose a fresh walklsp.rs:556-558)export biomelisting is O(n²) in sibling parsesexport.rs:34)gritlint-clippy(-D warnings, all targets) andgritlint-nextestgreen at66a753dforbidden-license.gritis compiled in the engineexpected/*.gritplugins compile in gritlint's engine with R36'sregister_diagnosticstub (8/8Ok). Loading them in Biome is R13npm/gritlint/README.md:107-108documents it, and the server announces the ROOT it serves (lsp.rs:517-525)publishdrops the event on a non-target errorexport-constructalso covers a body that does not parsecheckcompiled the bodyminorintent tells consumers about their own rulesinitializeis untestedservereturnsEnding::Refusedafterrefusewhateveruntil_exitreturns (lsp.rs:70-74,:120-130). Over stdio,-32803followed by closing stdin exits 2scan_contract.rs:814asserts findings throughdecide::plananddecide::scan, the callscheckmakes. The CLI adds only renderinginitialize_finishpanics if a memory client drops betweeninitializeandinitializedConnection::stdio()(apps/gritlint/src/main.rs:145). Over stdio, a client that leaves beforeinitializegives exit 1, and one that closes its read end mid-session gives exit 1 (probes)bitflagsmajorcargo denyprintedadvisories ok, bans ok, licenses ok, sources ok, and every vendor and dependency derivation builtapps/gritlint-e2e/srccontains no non-ASCII byte, so the two encodings give the same columns.lsp_contract.rs:367pins the encoding mapping (AE14)publishsends each URI the one set computed for its path (lsp.rs:360-366), against the last-edited buffer (:371-376)files_selectedlsp.rsreads no selection count, andchecknever restrictsdomain.rssizeParamNameRefusal/RelPathRefusalplacement, not as a findinglspis never run end to end with--configor--rules--rulesand given--config alt.json, loads and publishesunlicensed. The loader ischeck's (shell::load_workspace, KTD47)didChangeevents, load 32-36 on 32 cores). One openpackage.jsontakes 37 ms, 10 take 57 ms, and all 39packages/*/package.jsontake 99 ms, under the 150 ms of a fullcheckFilesdouble would pin wiring, not behaviour. The behaviour is latency, measured in R2, R3 and R24lsp_contract.rs:648pins the missing-ROOT case, the same empty setscheckgives when it selects nothing. Probe: a ROOT unreadable to the running user makescheckexit 2 ("selected no files"), while the server announces that ROOT and publishes an empty setgritlint-agent-journeydrives the real binary over stdio (apps/gritlint-e2e/src/lsp.rs:18), green in CI. In every probe,shutdownthenexitwith stdin still open exits 0RuleId::scope(domain.rs:630-632,export.rs:72-76), andexport-libraryrefuses any sibling callmain.rs:145opens it, andlsp.rsimports noshell(CONST-B4 comply in both rounds)target-not-scannedis never exercisedlog()makescheckexit 2 witha/package.json: target-not-scanned: seen, and the server publishestarget-not-scannedon the open document. No test constructs it. No file undercrates/gritlint_core/testsorapps/gritlint-e2e/srcnamestarget-not-scanned,NotScannedorTargetNotScanned. The only producer is an engine log ofDiagnosticKind::Other(decide.rs:453-456,engine.rs:104-109), mapped byGritlintError::target(error.rs:191-193). Red not shownlsp.rs:463-473)initializeis untestedinitializeexits 1 (lsp.rs:60-62)didChangewith emptycontentChangesis ignoredlsp.rs:105), so an empty array carries no text. Probe: no republish, andshutdownthenexitgives 0a/x.json, thena/y.jsononcexreads, the same one-at-a-time order ascheck(cannot read ./a/x.json). Across two batches, each open document names its own file in one eventcheckexits 2. A closed file's findings are out of scope in the same way (R31, KTD45)a/locked.json: target-unreadable: Permission denied (os error 13), andcheckexits 2 withcannot read. The conversion matches everyGritlintError::Iobelow ROOT with noErrorKindbranch (lsp.rs:427-433), the armlsp_contract.rs:607pinsReview receipt detail (from the reviewers' returns)
Both rounds ran on ce-code-review's full path (
depth: full). All five declared packs loaded from the02bf9653pin: boundary-testing, cell-architecture, mutation-discipline, package-topology and schema-laws, 41 rule files. Each round's learnings reviewer read every rule, and its project-standards reviewer treated all 16 constitution laws. Rules that apply, with the HEAD line each reviewer quoted:1eedb94)boundary-testing/no-mocks-on-internal-gluecrates/gritlint_core/tests/lsp_contract.rs:52,crates/gritlint_core/src/lsp.rs:541eedb94)boundary-testing/port-fakes-pass-the-port-contractcrates/gritlint_core/src/lsp.rs:461eedb94)boundary-testing/real-system-oraclesapps/gritlint-e2e/src/lsp.rs:17,apps/gritlint-e2e/src/lsp.rs:1751eedb94)boundary-testing/staged-protocol-evidencecrates/gritlint_core/src/lsp.rs:29,crates/gritlint_core/src/error.rs:1221eedb94)mutation-discipline/kill-survivors-with-contract-propertiescrates/gritlint_core/tests/export_contract.rs:142,crates/gritlint_core/tests/export_contract.rs:1861eedb94)mutation-discipline/suppress-by-shape-in-an-ignorer.cargo/mutants.toml:3,.cargo/mutants.toml:91eedb94)mutation-discipline/unkillable-survivors-are-dead-code.cargo/mutants.toml:131eedb94)package-topology/surface-changes-are-versioned.changeset/gritlint-lsp-biome-export.md:266a753d)boundary-testing/no-mocks-on-internal-gluecrates/gritlint_core/tests/lsp_contract.rs:61166a753d)boundary-testing/real-system-oraclescrates/gritlint_core/tests/lsp_contract.rs:63766a753d)cell-architecture/decode-never-castcrates/gritlint_core/src/lsp.rs:42766a753d)cell-architecture/four-channel-contractscrates/gritlint_core/src/error.rs:10166a753d)mutation-discipline/kill-survivors-with-contract-propertiescrates/gritlint_core/tests/lsp_contract.rs:62766a753d)package-topology/surface-changes-are-versioned.changeset/gritlint-lsp-biome-export.md:2(minor),:5(the summary; the reviewer's return cited it as:4)66a753d)schema-laws/refusals-beside-generated-lawscrates/gritlint_core/tests/lsp_contract.rs:60766a753d)schema-laws/tagged-unions-over-state-by-presencecrates/gritlint_core/src/error.rs:131Round 5 marked 8 rules as applying, 6 comply and 2 judged not applicable on reading; the delta marked 8, all comply. Every other rule file was judged not applicable, each with a stated reason in the reviewer's return.
1eedb94)66a753d)crates/gritlint_core/src/workspace.rs:27,crates/gritlint_core/src/domain.rs:36crates/gritlint_core/src/error.rs:101,crates/gritlint_core/src/domain.rs:829export-constructrow):crates/gritlint_core/src/export.rs:43,crates/gritlint_core/src/domain.rs:12crates/gritlint_core/src/error.rs:123,crates/gritlint_core/src/error.rs:131crates/gritlint_core/src/lsp.rs:219,crates/gritlint_core/src/lsp.rs:208crates/gritlint_core/src/domain.rs:827,crates/gritlint_core/src/decide.rs:49crates/gritlint_core/src/lsp.rs:46,apps/gritlint/src/main.rs:146crates/gritlint_core/src/lsp.rs:426,crates/gritlint_core/src/lsp.rs:545crates/gritlint_core/src/shell.rs:8,crates/gritlint_core/src/shell.rs:119crates/gritlint_core/src/lsp.rs:2,crates/gritlint_core/src/shell.rs:128crates/gritlint_core/src/export.rs:29,crates/gritlint_core/src/workspace.rs:81crates/gritlint_core/src/lsp.rs:545,crates/gritlint_core/tests/lsp_contract.rs:607crates/gritlint_core/src/workspace.rs:81,crates/gritlint_core/src/shell.rs:93crates/gritlint_core/src/lsp.rs:407,crates/gritlint_core/src/lsp.rs:545crates/gritlint_core/src/domain.rs:692,crates/gritlint_core/src/engine.rs:175crates/gritlint_core/src/lsp.rs:402,crates/gritlint_core/src/lsp.rs:427docs/plans/2026-10-12-2025-feat-gritlint-standalone-plan.md:263,crates/gritlint_core/src/lsp.rs:46crates/gritlint_core/src/error.rs:195,crates/gritlint_core/src/lsp.rs:428crates/gritlint_core/src/encode.rs:281,crates/gritlint_core/tests/common/mod.rs:12crates/gritlint_core/src/decide.rs:48,crates/gritlint_core/src/decide.rs:377apps/gritlint-e2e/src/export.rs:111,apps/gritlint-e2e/src/main.rs:142crates/gritlint_core/tests/lsp_contract.rs:611,crates/gritlint_core/tests/export_contract.rs:65apps/gritlint-e2e/src/export.rs:34,crates/gritlint_core/tests/export_contract.rs:186crates/gritlint_core/tests/lsp_contract.rs:627,crates/gritlint_core/tests/fixtures/export/expected/backslash-separator.grit:10crates/gritlint_core/tests/export_contract.rs:82,crates/gritlint_core/tests/lsp_contract.rs:626crates/gritlint_core/tests/lsp_contract.rs:632,crates/gritlint_core/tests/lsp_contract.rs:637crates/gritlint_core/src/domain.rs:691,crates/gritlint_core/src/export.rs:76crates/gritlint_core/src/lsp.rs:427,crates/gritlint_core/src/lsp.rs:428apps/gritlint/src/main.rs:126,apps/gritlint/src/main.rs:129crates/gritlint_core/tests/export_contract.rs:98,crates/gritlint_core/tests/export_contract.rs:100apps/gritlint-e2e/src/main.rs:143,crates/gritlint_core/tests/lsp_contract.rs:552crates/gritlint_core/tests/scan_contract.rs:181,crates/gritlint_core/tests/fixtures/export/expected/listing.txt:2Every reviewer in both rounds returned, and none timed out. The delta round's cross-model peer finished in 915 s of its 1,210 s deadline. Served models, counted as assistant turns in each reviewer's session transcript (no Grok or xai in any):
openrouter/deepseek/deepseek-v4.1-flash71,minimax-code/MiniMax-M3.1-Flash-Preview38openrouter/deepseek/deepseek-v4.1-flash49,minimax-code/MiniMax-M3.1-Flash-Preview4openrouter/deepseek/deepseek-v4.1-flash44,minimax-code/MiniMax-M3.1-Flash-Preview4openrouter/deepseek/deepseek-v4.1-flash32openrouter/deepseek/deepseek-v4.1-flash63,minimax-code/MiniMax-M3.1-Flash-Preview4openrouter/deepseek/deepseek-v4.1-flash77,minimax-code/MiniMax-M3.1-Flash-Preview4openrouter/deepseek/deepseek-v4.1-flash48,minimax-code/MiniMax-M3.1-Flash-Preview3openrouter/deepseek/deepseek-v4.1-flash35,minimax-code/MiniMax-M3.1-Flash-Preview5openrouter/deepseek/deepseek-v4.1-flash59,minimax-code/MiniMax-M3.1-Flash-Preview4openrouter/deepseek/deepseek-v4.1-flash51,minimax-code/MiniMax-M3.1-Flash-Preview13openrouter/deepseek/deepseek-v4.1-flash25,minimax-code/MiniMax-M3.1-Flash-Preview9openrouter/deepseek/deepseek-v4.1-flash34,minimax-code/MiniMax-M3.1-Flash-Preview7openrouter/deepseek/deepseek-v4.1-flash58openrouter/deepseek/deepseek-v4.1-flash10,minimax-code/MiniMax-M3.1-Flash-Preview8openrouter/deepseek/deepseek-v4.1-flash26,minimax-code/MiniMax-M3.1-Flash-Preview9openrouter/deepseek/deepseek-v4.1-flash31,minimax-code/MiniMax-M3.1-Flash-Preview9openrouter/deepseek/deepseek-v4.1-flash43,minimax-code/MiniMax-M3.1-Flash-Preview9minimax-code/MiniMax-M3.1-Flash-Preview14,openrouter/deepseek/deepseek-v4.1-flash10The delta round's merge and report leaves ran on
anthropic/claude-haiku-4-5. They synthesize the reviewers' returns and do not review.Cross-model passes on the 2026-10-13 body updates. Both used an
adversarial-revieweron DeepSeek, worked read-only, and finished without timing out; neither used Grok or xai.openrouter/deepseek/deepseek-v4.1-flash, 43 turns, 3m53s) checked the CI, ci-wall-clock and residual sections against the jobs API, the job logs, the66a753dtree and the pack at763fd663. It filed 2 P3 findings, both applied:flake-outputs-build-once-on-mainallows two no-save cases, not one, and a receipt row cited the changeset's blank line 4.openrouter/deepseek/deepseek-v4.1-flash, 16 turns, 1m59s) checked the whole body for red evidence that is not allowed. It found none. It confirmed the unmodified-1eedb94red by rerunning both binaries, theexport.rsbyte-identity, and the reading-based gaps. It filed 1 P3, applied: the red contradicts the test's diagnostics assertion (:623-629), not its first one.ce-compound
Skipped. The durable reasoning is in the plan (KTD43 to KTD55), the contract tests (
lsp_contract.rs,export_contract.rs,scan_contract.rs) and this body. No lesson from this unit passes the counterfactual bar.Mutants (
cargo mutants --workspace --list --in-diff, no mutation run)186 of the workspace's 627 mutant ids fall in this PR's diff: lsp.rs 84, export.rs 34, decide.rs 13, workspace.rs 13, error.rs 12, domain.rs 9, engine.rs 9, shell.rs 8, encode.rs 4.
main's Mutants workflow measures them after merge.ids
CI wall clock (jobs API): x86_64 12m05s against the 10-minute PR target
Target: every pull-request job finishes within 10 minutes (plan R27; ci-wall-clock
pr-gate-runs-only-what-the-diff-affects). On run 38307881354 at66a753d,Nix · x86_64-linuxtook 12m05s, 2m05s over. Every other job finished under the target. Budgets are measured, not enforced, sotimeout-minutesstays at the uniform 60 (KTD53).x86_64, step by step (
started_at/completed_atof each step fromGET /repos/systemfsoftware/gritlint/actions/runs/38307881354/jobs):outPathsnix-store-x86_64-linux-88683b6d…. The prefix hit was this PR's own entry…f219aea9…(2.30 GB), saved at 20:04:56Z by run 38300038684 at30af161, before #17 stopped pull-request savesnix copyfrom the restored bundleInside step 9, from the derivation log prefixes (UTC):
gritlint-deps5m05s,gritlint-linux-x64-deps4m55s, andgritlint-linux-arm64-deps(cross) 3m29s (21:06:25-21:09:54).gritlint2m21s,gritlint-linux-x642m21s,gritlint-nextest1m30s, thegritlint-benchbinary 1m11s, andgritlint-linux-arm64(cross) 1m10s.checks.x86_64-linux.gritlint-bench, the PR3 corpus benchmark, runs as the last derivation for 1m19s.Cause: structural, not a cold cache after #17. The five crane dependency derivations take
Cargo.lockas input, and this PR changes it:lsp-serverandlsp-typesadd 68 lines againstmain, and0c45d15gaveapps/gritlintits ownlsp-serveredge. No cache entry on any ref holds these derivations:main's bundle (…76f3d330…,c82913f) predates the new crates.…f219aea9…) holds30af161's lockfile.So every run of this PR since
0c45d15rebuilds them. Run 38304323655 at1eedb94had the sameCargo.lockas66a753d(no dependency input changed between the two). It rebuilt the same five dependency derivations, identical store hashes, that run 38307881354 rebuilt, and took 10m18s. Two consecutive runs at the same inputs were both cold for dependencies. That pair is the evidence that a re-run cannot be warm. A re-run of 38307881354 restores the samef219aea9entry and rebuilds the same 31 derivations, so no warm re-run number exists to record. I also could not start one:gh run rerun 38307881354,POST /actions/runs/38307881354/rerun,POST /actions/jobs/115006212507/rerunand aworkflow_dispatchofci.ymleach return 403 "Resource not accessible by personal access token".Measured context (x86_64 and aarch64
Nixjobs, jobs API):66a753d1eedb94, sameCargo.lock30af161, pre-#17 rule (exported and saved)mainpush #18 (workspace versions)mainpush #19 (no cache)mainpush #16ci.ymland docs)mainpush #17Follow-up F3. Owner: the gritlint ci-wall-clock repin, which enrols the pack and is queued after this PR (not started here). Trigger: that repin's PR. Under
pr-gate-runs-only-what-the-diff-affects, it must bring a dependency-changing pull request's x86_64 job under 10 minutes or record the measured exception. Until then, each PR whoseCargo.lockdiffers frommain's newest bundle pays this rebuild. On this run that was about 5 minutes of critical path, plus 1m19s for the corpus benchmark at the tail. After this PR merges,main's push saves a bundle with the new crates, and later PRs that leaveCargo.lockalone restore it.Readiness
Head
66a753d8e9d5429a107f0b0280445d1cc40ebdc5, unchanged by the 2026-10-13 body update: CI run 38307881354 and Changeset Check 38307881969 green; delta receipt20261012-214525-6e25ba95Ready to merge; containsmainat0e2e07b. #19, the layer below, has merged, so this is the only open layer on trunkmain. Merging needs the one approving review themainruleset requires; the operator merges.